Eye Gaze and Attention Patterns Power a New AI Screen for Autism in Indian

Researchers in India developed a two-part AI screening system for autism spectrum disorder using eye gaze and attention patterns rather than questionnaire-based tools. In a study published in Complex & Intelligent Systems, the system reported accuracy above 97% on a dataset of Indian children. The approach targets subjectivity and potential cultural bias in parent-reported screening, such as the M-CHAT. The primary dataset included 14 autistic and 14 non-autistic children, balanced for training. A hybrid model combining ConvNeXtBase with LightGBM reached 97.65% accuracy and up to 98% ROC-AUC, with reliability supported by Cohen’s Kappa across multiple train-test splits and a paired t-test (t=14.4, p<0.001). A second task measured attention using histogram of oriented gradients and a linear SVM, showing notably lower attention and engagement in the autistic group during live activities.